baseclass(template)use subclass field+!vptr

In my baseclass template is able to use a subclass field. This reusable technique is reusable in a very restrictive context —

  1. the base and sub classes represent market data messages, used in a reinterpret_cast context, with zero padding. Therefore, a proposed vptr would add a pointer and mess up reinterpret_cast.
  2. multiple (say 11) subclasses have a “qty” field, so I don’t want code duplication 11 times
  3. The order among the fields is fixed in struct definition. Without this restriction, I would define the “price” field in a base struct and also a getPrice() method. With a subclass instance, CRTP could probably work like static_cast<Subclass const*>(ptr)->getPrice() but the “price” field’s physical offset would be be dictated by the compiler not according to my struct definition

My technique uses CRTP but no SFINAE no enable_if.

My technique is easier to /internalize/, as it relies on simple overload resolution + simple type deduction. In contrast, the SFINAE technique used in my RTS codebase is off-putting and alien to me

specialize class- but !!function- templates

A fundamental TMP technique is class template specialization.

Class templates’ flexibility can only be achieved via specialization but function templates’ flexibility can be achieved via overload ! Overload is much simpler than specialization.

In [[c++coding standard]], Alexandrescu/Sutter said “Don’t specialize func templates”, for multiple reasons.

After remembering this sound byte, it’s probably important to remember one of the reasons, for IV (halo) and zbs.

c++TMP^other QQ topics #java

Alexandrescu’s TMP techniques (not “designs”) are very tricky (not “complex”). They require absorbency, but do they enhance latency? Do they get you higher jobs with lower stress?

I need to make time-allocation decisions among QQ topics, including TMP

In terms of latency, Well, java can now rival c++ in latency. The technical reasons are not obvious nor intuitive, but not my focus today. Just an observed fact which discredits conventional wisdom and our assumptions.

— zbs, based on continued relevance :

TMP is needed when reaching next level in c++ zbs.

TMP is more time-honored than many c++0x features.

Many new c++0x features were added for TMP. I feel TMP is the main innovation front across c++ language n standard development. C++ lost many battles in the language war but no other languages offer anything close to TMP features.

— As QQ

Will C++TMP (and rvr) QQ turn out similar to java bytecode engineering, reflection, generics? (Even in such a scenario, TMP still offers better roti than Qz.) Actually TMP is quizzed more than those. The c++ guru interviewers often adore TMP.. cult following.

EJB is an add-on package .. different category, not an advanced core language feature.

When TMP is not quizzed you may still get opportunities to showcase your halo. Many interviewers ask open-ended questions.

TMP techniques would remain a halo for years to come. Classic QQ topic.

— GTD: templates are never needed in greenfield projects. Occasionally relevant in understanding existing code base such as etsflow, STL, boost..

Q: are there rare projects using TMP and offer me an opportunity to outshine others, gain GTD advantage ..?
A: I guess it’s one in 10 or 20. Here are some factors:

Within a given project codebase, TMP is a powerful tool for DRY improvement and re-usability , but such reusability  is over-rated in most projects.

DRY (don’t repeat yourself) is practiced more widely, but I feel TMP techniques replace 100 lines of code duplication with 20 lines of unreadable code.


MI loses type info about subclasses

[[Alexandrescu]] pointed out a fundamental weakness in MI for library design — base classes “do not have enough type information to carry out their tasks”, and “MI loses type information (about subclasses) which abounds in templates”

This statement can only be understood based on TMP. I like and will repeat this statement in QQ interviews. If interviewer is knowledgeable enough about TMP to quiz me further, I would say

“TMP can be combined with inheritance. I don’t remember the various TMP techniques that make use of the type information of subtypes”.

TMP is a compile-time technique so type information is more available.

template specialization based on NDTTP=true

We know it’s possible to specialize a template for a concrete type like int or std::string, but I didn’t know that It’s also possible to

… specialize a (class or function) template for a particular compile-time const value (like “true”) of a NDTTP (like “bool flag”)

  • On [[Alexandrescu]] Page xii , Scott Meyers showed an elegant example of specializing for “true”. Note “true” is a value, not a data type !
  • P 34 has a longer example.

Note on reading TMP code — the template specialization syntax is clumsy and can add noise to the signal. Better ignore the syntax rules for now to focus on the gist.

parts@mv-semantics impl: helicopter/hist view

Move semantics is 90% compile-time + 10% run-time programming. 95% in-the-fabric and invisible to us.

  • 90% TMP
    • 70% is about special casts — in the form of std::move and std::forward
    • std::swap
  • 10% traditional programming
    • RAII to destroy the “robbed” rvalue object
    • heap ptr assignment [1] to rob the resource

[1] This “stealing” is tip of the iceberg, the most visible part of move semantics. Bare-hand stealing was doable in c++03, but too dangerous. The rvr is a fundamental language feature to make the “steal” safer:

  • a safety device around the dangerous “steal”
  • a safety glove
  • a controlled destruction

The resource is typically a heapy thingy, common fixture in all STL containers and+ std::string + smart pointers. Therefore, move-semantics is widely used only in libraries, not in applications.

enable_shared_from_this #CRTP

  • std::enable_shared_from_this is a base class template
  • shared_from_this() is the member function provided
  • CRTP is needed when you derive from this base class.
  • underlying problem (target of this solution) — buggy design where two separate “clubs” centered around the same raw ptr. Each club thinks it owns the raw ptr and would destroy it independently.
  • usage example —
    • your host class is already managed via a shared_ptr
    • your instance method need to create new shared_ptr objects from “this”. Usually, you instantiate a new shared_ptr into an existing “club” by copy-construction from an existing shared_ptr but now you only have a raw ptr in “this”.
    • Note if you only need to access “this” you won’t need the complexity here.

Pimco asked it in 2017!

SCB-FM IV by architect #shared_ptr upcast

Q: how does the compiler accept this code:
shared_ptr<C> aa = myDerSharedPtr; //myDerSharedPtr is a shared_ptr<D> object

%%Q: shared_ptr<C> has a copy ctor and also a conversion ctor accepting a C raw ptr, but here we are passing in a shared_ptr<D> instance. How does compiler handle it?
%%A: I guess shared_ptr<D> has a conversion operator returning a D raw ptr, but this is not used.
AA: there’s a cvctor (conversion ctor) template<class U> shared_ptr(shared_ptr<U>…) — a TMP trick. shows cvctor is featured in the standard API.

A: P 176 [[moreEffC++]] uses OOC instead of cvctor. The OOC is a template member of the shared_ptr class template.

On a separate note, The github experiment also reveals — If a function lvr param is shared_ptr<C> & and you pass in a shared_ptr<D>, compiler will complain about assigning an rvalue (i.e. anonymous temp) object to an lvalue reference — a key insight into rvr + rvalue objects.

Q3: just when is the memory freed for temp objects like q[ string1 + string2 ]
%%A: at an unspecified time. A custom string implementation could use COW, in a single-threaded project. This is a common practice in many pre-c++11 libraries
A(from architect): after the semicolon

Q3b: how can you extend the lifetime of those naturally occurring temp object?
A: assign the temp to a “const ref” variable.

Q: what are your favorite c++11/14 features? See ## c++11 features I understand as significant

Q: OK you briefly mentioned move semantic..what is it?

struct C{ //tested
  virtual void f(){/*..*/}
  ~C(){     cout&lt;&lt;"C dtor\n";  } //non-virtual
struct D: public C{
  string s;
  D(): s("def"){}
  ~D(){     cout&lt;&lt;"D dtor\n";  }
D createD(){return D();} //return by value! probably via RVO
int main(){
  C const &amp; trade = createD();

Q: is string memory freed?
%%A: yes. Verified

Q: what if the string field is in D?
%%A: yes. Verified

I believe the temp D object is on stack and is guaranteed to be destructed. Since the D ctor called base ctor, the dtor sequence is guaranteed to be ~D then ~C.

::value^::type #typedef

This TMP technique is the simplest but not for the uninitiated.

In type_traits.h, many templates expose a ::value or ::type construct. Often these two “members” are the only visible output from the type trait meta-functions.

  • ::value is a static field, typically true/false
  • ::type is a member typedef, typically related to the type argument T, but can be “void” like in enable_if


TowerResearch IV: c++,algo,past prj..

Q1: how could you use enum to improve performance
AA: use it to replace strings
AA: enum is used in template programming to move computation from run time to compile time. See Q4 and SFINAE in github.
%%A: see clever use of enum(char?) in demanding c++ app

Q1b: you put in your change but it turns out to hurt performance. Any idea?

Q4: how do you compute Fib(N) like Fib(99999) at compile time using template programming?
A: See my Fibonacci code in github

Q: how would you use multicast to send a message to a group of registered users?
%%A: less data copying than TCP

Q: In O(1) space, Given an unsorted array of natural numbers, how do you find any pair to produce a target sum? What’s your best time complexity?
%%A: if I assume the integer size is 32-bit then a fixed-size O(1) space radix structure can sort the array in O(N). See also my blog post on [[locate a pair with targetSum=55 #bbg IV #Morris]]
%%A: if integer size is unlimited, then I think the best is O(NN)

Q: How is your muni bond pricing engine designed internally?
Q2b: what kind of pricing rules live in the cache?
Q2c: how many threads to react to a given event?

Q: how did you calculate FX fwd points?
%%A: use the interest rates in the two currencies and the spot FX rate.

Q: how is your implied volatility computed from an option price?
%%A: if there’s not a closed-form formula, then Newton’s method or something similar should be able to converge quickly.

Q3 (OO design): How would you design a game software with Animal objects, various birds, and fish?
Q3a: Say the system is running fine, but now you need to add ostrich class?

template template param

Andrei Alexandrescu briefly introduced this technique .. I think he said it is one of the most powerful TMP techniques.

Vague recollection — one of the template parameters is itself a template.

One common feature of templ-templ param — the param P “template<typename>P” is usually used as “P<S>”, where S is a sister parameter to P ! has a good tutorial. shows operator<<(). I have it in my github

The variadic version is broken because the template tries to accept any class template having 1 or more parameters in the definition. This includes std::basic_string, which already has an operator<<() that conflicts with the variadic version.

This also helps me understand that the 2-param version (using templ-templ) is really a template that accepts any class-template having two type-params by definition. Note the two type-params can use default values.

— See my github demo

First and possibly biggest challenge is — understand the limitation of traditional solutions. I would say this is 90% of the struggle for a novice.


LiquidNet IV c++

Q1: partition list into good^bad sections

Q3b: fix a sample source code that passes unique_ptr<int> by value to a foo() function:

unique_ptr<int> foo(unique_ptr<int> kk)

Q3b.1: give examples of move semantic

Q3b.2: write a template solution to check if the template argument T is a pointer type. I don’t have a good solution but I mentioned SFINAE. My github has a sample sfinae code to check for specific types of pointers. In contrast, std::is_pointer doesn’t handle ptr to member etc.

How about static_assert? Yes tested

See post on template type arg validation

Q3a: improve

string & find(list<Person> li, string name){
  for (auto i=li.begin(); i!=li.end(); i++)
    if (*i == name) return i->nickname;
  return "";

%%A: last return will crash due to return-by-ref. If my function’s return value can be empty, then I usually return by pointer
%%A: equality check is dubious. Better be explicit like Person(name) or i->convertToString()
%%A: li.end() can be cached
%%A: ++i
%%A: li and name should really be const ref

Q4: predict output
%%A: I feel #2/3 are bad code but compiler may not be able to

struct B{void foo(){cout<<"hello\n";}; b = new B(); b->foo(); //#1
delete b;
b->foo(); //#2 .. what?
b->foo(); //#3 .. what?

favor std::begin(arrayOrContainer) explains some important details.

Q: So how do we choose between

  • this free global function
  • the container member function cont::begin() / end()?

%%A: Basically, I would always use std::begin() instead of cont.begin() esp. in template-enable programs.

another basic difference iterator^container^algo

(Why bother? Well, you need to know these when you debug STL or extend STL.)

– Containers — are class templates. 100%
– Algorithms — are function templates. 100%

– iterators? A beginner can safely assume that Most of the time iterators are defined inside each container template as a Member type. Since a container has a dummy type T, the iterator must be a class template of T (or a typedef thereof), a Nested class template of T, presumably, inferred from the syntax :


– The container/algo/iterator adapters are typically class templates
– functor are typically class templates

A trivial consequence — the declarations of containers and iterators are complicated by the templates. Algorithm declarations are simpler in comparison.